Rahul Baxi

I build tools that help people share what they create, and frameworks that help us trust the AI behind them.

I'm the Cofounder & CTO of Smartypans, where we're reinventing how cooking becomes content. Our sensor-AI cookware records cooking in real time and automates recipes, nutrition, and content workflows, so creators can focus on creativity, not admin work.

I'm also an independent AI safety and alignment researcher. My work focuses on measuring robustness in frontier language models across semantic, epistemic, and ethical dimensions. I developed CDCT(Compression-Decay Comprehension Test), which reveals how models lose meaning under constraint pressure, and DDFT(Drill-Down and Fabricate Test), which probes epistemic integrity: whether a model truly knows something or fabricates plausible-sounding detail. Together, they show that robustness dimensions are orthogonal. A model that handles compression well can still hallucinate under drill-down, and vice versa. This has direct implications for how we evaluate and certify AI systems before deployment.

My research is grounded in a practical problem. At Smartypans, AI failures in cooking guidance can cause real harm. That experience shapes my broader interest in alignment as comprehension, not just compliance, drawing on Vedic epistemological traditions to bridge ancient frameworks for knowledge integrity with modern AI safety.